← All updates

Genie 4 Goes Live

Plus: GPT-Live 1.1 reaches the API, Muse Wonder leads legal benchmark

Good Morning!

Google opened its AI world generator to developers. OpenAI opened its new voice model. Apparently, the future is something you can both walk through and talk to.

DIT
Model Ranking
THURSDAY
Top 10 text models by real-world token usage · Nov 19, 2026 (UTC)
Real-world use:
Agent
Knowledge
Chat
Model
tokens
WoW (11/12)
1
Hy4 flash
· Tencent
22.64T
NEW
2
DeepSeek V4.1 Flash
· DeepSeek
11.77T
▼1
+0%
3
GPT-6 Luna
· OpenAI
1.60T
▲1
+18%
4
GLM 5.3 Flash
· Z.ai
1.47T
▼1
-11%
5
Hy4
· Tencent
1.44T
▼3
-47%
6
Claude Opus 5.5
· Anthropic
1.08T
▼1
-9%
7
DeepSeek V4 Flash
· DeepSeek
0.68T
▼1
-20%
8
MiniMax M3.1 Flash
· MiniMax
0.67T
▲6
+57%
9
MiMo-V2.6-Flash
· Xiaomi
0.60T
▼1
-5%
10
Nemotron 3 Ultra
· NVIDIA
0.56T
▲2
+24%
How we rank: tokens processed over the full 24 hours of Nov 19, 2026 (00:00–24:00 UTC) on OpenRouter. Usage lanes follow OpenRouter's task classification, grouped by DIT Radar into Agent / Knowledge / Chat and distributed to each model by our estimate.
Source: OpenRouter. The ranking reflects DIT Radar's interpretation of the data.

Genie 4 Opens Developer API

Google Genie 4 world-model artwork
Image: Google

TL;DR: Following its Project Genie rollout, Google made Genie 4 available to developers through the Gemini API's Interactions API and Google AI Studio. Developers can generate continuous 720p worlds at 24fps with synchronized audio, navigate with keyboard controls and edit scenes through live prompts. Text prompts, reference images or both can start a world. Read more →


GPT-Live 1.1 Reaches the API

OpenAI GPT-Live documentation artwork
Image: OpenAI

TL;DR: OpenAI opened GPT-Live-1.1 to API developers after using it in ChatGPT voice since October 15. The full-duplex model listens while speaking, improves transcription and keeps accents more consistent across languages. Active sessions cost $0.05 per minute; paused WebRTC standby costs $0.02 per minute. Backend models and tool use are billed separately. Read more →


Muse Wonder Leads Legora BAR

Meta headquarters in Menlo Park, California
Photo: Reuters / SIPA USA

TL;DR: Muse Wonder ranked highest on overall quality in Legora's November BAR evaluation of legal work, leading analysis and answering, extraction, drafting and editing. GPT-6.2 Astra led comparison and review tasks, while Sol delivered above-average quality at lower cost. The results compare models inside Legora's legal workflow and evaluation setup. Read more →


Together Adds Free GPU Caching

Together Storage Accelerator announcement artwork
Image: Together AI

TL;DR: Together launched Storage Accelerator, a caching layer included with its GPU Clusters. It keeps downloaded models and datasets on cluster NVMe so nodes can reuse them without repeatedly fetching from object storage or Hugging Face. Together measured roughly 20 times faster warm reads on an H100 node for a 2 GiB object from Cloudflare R2. Read more →


Gemini Loads Tools on Demand

Gemini symbol against a dark blue background
Image: Google

TL;DR: Google previewed Tool Retrieval in the Gemini Interactions API. Agents can see tool names and descriptions while loading full parameter schemas only when needed, reducing the overhead of large tool libraries. The feature supports Gemini 3.6 and later, including Gemini 4. It has no separate charge; standard token billing still applies. Read more →


Meta Teaches Agents Better Memory

Meta research chart comparing optimized agent memory with hand-written memory strategies across three tasks
Image: Meta / RAM research team

TL;DR: Meta researchers used Kimi-K2.6 to learn better ways for agents to save, organize and retrieve memory. On three Continual Learning Bench tasks, optimized memory prompts averaged 56.0 versus 44.6 for the best hand-written prompts. Distilling those habits into the model then raised its score under generic prompts from 37.6 to 43.1. Read more →


Ai2 Opens Theorizer for Research

Ai2 Theorizer graphic showing literature search, evidence synthesis and hypothesis generation in Asta
Image: Ai2

TL;DR: Ai2 launched Theorizer inside Asta and released its code. The research agent combines findings from scientific papers to propose explanations, linking each idea to supporting evidence and where it may apply. Researchers can try it through Asta with a limited token allowance or run the open library locally using paid model APIs. Read more →


Top AI models. Lower costs.

Access GPT, Claude, Gemini, and more through one OpenAI-compatible API—and save up to 70% on AI API costs. Create your DIT account, get your API key, and start building.